the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 10 Mar 2026
GeoSIRR 1.0: Conversational Geological Cross-Section Modeling Using Large Language Models
Abstract. Geological cross-sections are a fundamental tool for subsurface interpretation, yet their construction remains a labor-intensive and largely manual process that relies on expert judgment and structured numerical inputs. While recent advances in artificial intelligence have enhanced specific geoscientific workflows, no existing method enables the direct generation and iterative refinement of geological cross-sections from unstructured natural language descriptions. In this paper, we present GeoSIRR 1.0 (Geological Section Interpretation, Reconstruction & Refinement), a novel modeling framework that leverages large language models (LLMs) to translate free-form geological narratives into structured, coordinate-based cross-section geometries. GeoSIRR introduces a domain-specific language (DSL) for representing geological bodies as topologically consistent polygons and integrates automated geometric and geological validation to ensure continuity, stratigraphic consistency, and structural plausibility. The framework supports both initial model generation and conversational refinement, allowing users to iteratively modify cross-sections using natural language commands while preserving existing geometry. We demonstrate the capabilities of GeoSIRR through multiple geological scenarios, including faulted sedimentary systems, intrusive bodies, and progradational deltaic sequences, and assess repeatability across multiple generation runs. Results show that GeoSIRR consistently produces geologically plausible cross-sections and effectively incorporates conceptual refinements with reduced generation time compared to initial model construction. By directly linking qualitative geological reasoning with quantitative geometric modeling, GeoSIRR provides a self-contained, dialogue-driven approach to cross-section construction that complements existing modeling tools and offers new opportunities for education, exploratory analysis, and rapid scenario development in subsurface geoscience.
- Preprint
(1786 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2025-6545', Anonymous Referee #1, 10 Jul 2026
-
AC2: 'Reply on RC1', Denis Anikiev, 26 Aug 2026
We thank the referee 1 for the detailed and critical assessment. The concerns raised have helped us to rethink both the practical accessibility of the framework and the evidence required to establish its scientific contribution. We agree that the original manuscript did not sufficiently discuss access to alternative LLMs, performance for more complex geological configurations, or the repeatability of the generated geometries. The comment has also helped us to distinguish generation repeatability more carefully from geological uncertainty in a real subsurface interpretation. We have addressed these issues as described below.
Reproducibility and access to LLMs
> However, GeoSIRR 1.0 in its current form depends on GPT language models, and an OpenAI API key is required to run the cross-section reconstructions. When attempting to run the provided examples, I received error messages indicating that my API quota had been exceeded and that additional credits would need to be purchased to proceed. This dependency renders GeoSIRR 1.0 effectively inaccessible for reviewing or users without an active paid API subscription, meaning the results cannot be freely tested or reproduced.
We thank the referee for this observation and agree that dependence on a paid API was an important practical limitation of the original version. GeoSIRR 1.0.1 is already available through Zenodo:
Anikiev, D., & Mosquera, J. E. (2026). GeoSIRR: Geological Section Interpretation, Reconstruction & Refinement (Version 1.0.1) [Computer software]. https://doi.org/10.5281/zenodo.22097054
as well as GitHub:
https://github.com/CPG-KFUPM/GeoSIRR
In this version, we have extended the framework to support both OpenAI and Ollama backends. Ollama allows a model installed on local hardware to be used without an OpenAI API key or paid OpenAI credits.Moreover, we have deployed GeoSIRR at https://geosirr-studio.onrender.com, which provides a hosted graphical web service for generating and refining cross-sections based on the original GeoSIRR workflow. During the review process, this hosted service is configured so that referees can generate and refine cross-sections without registration or providing their own API key. We will maintain this access until the review process is complete.
Pre-generated GeoSIRR text definitions can also be inspected, validated, and rendered using either GeoSIRR 1.0.0 or 1.0.1 without repeating an LLM call. Thus, the geometric results can be examined independently of access to an LLM provider, although reproducing the stochastic generation process still requires LLM calls. In the revised manuscript, we will clarify this distinction and explain that use of a local model does not guarantee the same output quality, runtime, or geological interpretation as another model, because these aspects depend on the selected model, its configuration, and the available hardware.
More complex geological scenario
> Furthermore, the manuscript presents only relatively simple tectonic geometries as examples and does not thoroughly demonstrate how the method performs on more complex geometries or tectonic relationships. The authors indicate that increased complexity leads to a significant increase in computation time, yet no further testing is provided for geologically meaningful complex scenarios.
We thank the referee for this constructive suggestion and agree that a more geologically involved example would strengthen the manuscript. We plan to add a classical extensional-tectonics example containing several interacting structural elements and sedimentary basin fill. We will use this example to test whether GeoSIRR can translate a more complex geological description into a valid geometry while preserving the relationships among the interacting structures and sedimentary bodies. We will also examine how this increase in geological complexity affects generation time. We will also add a repeated-generation experiment for this scenario and report summary measures of validity, generation time, and run-to-run geometric variation. These experiments will provide evidence of both the framework's representational capability and the aspects of the generated geometry that remain sensitive to stochastic generation.
Repeatability and generated-geometry variability> Moreover, the limitations related to GeoSIRR 1.0, which are noted in the manuscript (no uncertainty quantification, unclear generation performance and repeatability), are serious and the model should not be considered without these issues being properly assessed.
We agree that generation performance and repeatability need to be assessed explicitly. In the revised manuscript, we will add repeated-generation experiments using the same model and workflow for a baseline description and a more constrained variant. We will track the validation success, generation effort, and run-to-run geometric variation, supported by numerical summaries and visualizations of the generated geometries. We believe that this comparison will help demonstrate the reproducibility of the workflow, show how prompt specificity affects geometric variability, and clarify that variability among generated models is not equivalent to geological uncertainty in a real subsurface interpretation. We thank the referee for prompting us to provide this evidence.
Citation: https://doi.org/10.5194/egusphere-2025-6545-AC2
-
AC2: 'Reply on RC1', Denis Anikiev, 26 Aug 2026
-
RC2: 'Comment on egusphere-2025-6545', Anonymous Referee #2, 30 Jul 2026
The manuscript introduces GeoSIRR 1.0, a framework for translating natural language geological descriptions into 2D coordinate-based cross-sections using Large Language Models. The use of a Domain-Specific Language to bridge qualitative geological narratives with quantitative geometric modeling is a valuable contribution. The manuscript is generally well written, and the presented geological examples effectively demonstrate the capabilities of the framework. I only have a few minor comments that should be addressed before publication.
- The manuscript reports varying generation times, with the laccolith intrusion example requiring up to 956.6 seconds. It would be helpful if the authors could briefly clarify whether this latency is primarily due to LLM API inference or to internal processing within the GeoSIRR workflow.
- In Section 4.4.3, topological consistency is verified by comparing the total polygon area with the area of the domain rectangle. While this is an efficient approach, it may, in theory, miss cases where overlap is balanced by a corresponding gap elsewhere. A brief discussion of this limitation and its possible improvement in future work would strengthen the discussion.
- In Example 2, the authors note that contradictory input constraints (layer thicknesses exceeding the specified section depth) resulted in different solutions across repeated runs. Since the workflow includes an agent to validate user prompts, it would be useful to briefly explain why this inconsistency was not detected before generation and to provide a simple guideline for users when specifying section dimensions.
- The current method demonstrated using GPT-5 only. Although this is sufficient to demonstrate the framework, a brief discussion of future benchmarking against other LLMs (e.g., open-weight or domain-specific models) would broaden the work's impact and help evaluate trade-offs in accuracy, generation time, and computational cost.
- The case studies demonstrate the framework well; however, the Results section would benefit from a small summary table listing key metrics for each example, such as generation time, number of vertices/polygons, validation status, and refinement time. This would make it easier for readers to compare the examples and assess the system's overall performance.
- The manuscript would benefit from careful proofreading to eliminate a few typographical and grammatical errors. For example, "interpet" should be corrected to "interpret" (p. 3), and "engeneering" should be corrected to "engineering" (p. 9).
Citation: https://doi.org/10.5194/egusphere-2025-6545-RC2 -
AC3: 'Reply on RC2', Denis Anikiev, 26 Aug 2026
We thank the referee 2 for the positive evaluation and for the helpful suggestions. We address each point below.
Generation latency
> The manuscript reports varying generation times, with the laccolith intrusion example requiring up to 956.6 seconds. It would be helpful if the authors could briefly clarify whether this latency is primarily due to LLM API inference or to internal processing within the GeoSIRR workflow.Thanks for the comment. The reported 956.6 s is an end-to-end wall-clock value, not the runtime of the geometric checks alone. The original measurement unfortunately did not separate the LLM calls from individual local operations, so we cannot reconstruct a precise stage-by-stage division for that run and will not attribute its duration to a specific cause that was not measured.We have added stage-level timing in GeoSIRR 1.0.1:Anikiev, D., & Mosquera, J. E. (2026). GeoSIRR: Geological Section Interpretation, Reconstruction & Refinement (Version 1.0.1) [Computer software]. https://doi.org/10.5281/zenodo.22097054
We will summarize these aspects also in the revised manuscript. This will distinguish time spent on LLM-based operations from time spent on local validation and rendering, and will clarify which stages dominate the reported end-to-end latency.Topological consistency check> In Section 4.4.3, topological consistency is verified by comparing the total polygon area with the area of the domain rectangle. While this is an efficient approach, it may, in theory, miss cases where overlap is balanced by a corresponding gap elsewhere. A brief discussion of this limitation and its possible improvement in future work would strengthen the discussion.We thank the referee for drawing attention to this point. The description in the manuscript was indeed incomplete. The current validation procedure does not rely only on comparison of the total polygon area with the domain area. It also checks the validity of individual polygons, overlaps between polygons, gaps within the rectangular domain, geometry extending beyond the domain, and whether the full set of polygons forms one continuous rectangular section. We will revise the methodological description to explain these checks more accurately, and we will add the requested limitation discussion. Also, we would like to emphasize that the validation remains a two-dimensional geometric check performed with a numerical tolerance. It cannot replace geological interpretation or guarantee that a geometrically valid cross-section is consistent with a full kinematic history or regional geological setting. These are important directions for future development.Contradictory constraints in the laccolith example> In Example 2, the authors note that contradictory input constraints (layer thicknesses exceeding the specified section depth) resulted in different solutions across repeated runs. Since the workflow includes an agent to validate user prompts, it would be useful to briefly explain why this inconsistency was not detected before generation and to provide a simple guideline for users when specifying section dimensions.We thank the referee for this observation. The clarification stage screens whether the description contains sufficient and unambiguous geological information, such as the number of layers and structural setting. It does not sum all numerical constraints. In the laccolith example, the individual dimensions were present, so the description was considered complete, but their combined vertical extent was inconsistent with the declared section depth. We will clarify this distinction in the revised manuscript and add a practical guideline for users. When layer thicknesses, basement thickness, intrusion dimensions, air units, or other vertical dimensions are specified, their combined requirements should be compatible with the total depth of the section. Contradictory fixed constraints should be resolved before generation. We think that a more deterministic numerical consistency check would be a useful future extension of the framework.Benchmarking other LLMs> The current method demonstrated using GPT-5 only. Although this is sufficient to demonstrate the framework, a brief discussion of future benchmarking against other LLMs (e.g., open-weight or domain-specific models) would broaden the work's impact and help evaluate trade-offs in accuracy, generation time, and computational cost.We agree and thank the referee for this suggestion. The original examples use GPT-5, whereas the new repeated-generation experiments that we plan to include in the revised manuscript will use an open-weight gemma4:31b model. The new version of the framework (GeoSIRR 1.0.1) supports both backends, broadening access beyond OpenAI. In the revised manuscript we will identify the model and backend for each result and will discuss the need for a controlled comparison of general-purpose, open-weight, and domain-specific models. Such a benchmark should evaluate geological consistency, prompt compliance, repeatability, generation time and hardware requirements. The preliminary experiments already demonstrated backend portability, but we don't plan to present them as a controlled cross-model benchmark as this is beyond the scope of the current work.Summary table for the examples> The case studies demonstrate the framework well; however, the Results section would benefit from a small summary table listing key metrics for each example, such as generation time, number of vertices/polygons, validation status, and refinement time. This would make it easier for readers to compare the examples and assess the system's overall performance.We agree and will add a summary table to the Results section. For each initial example and refinement, the table will list the model and backend, generation or refinement time, number of vertices, and number of polygons. We thank the referee for this suggestion, which will help readers to compare the examples and assess the framework's performance more easily.> The manuscript would benefit from careful proofreading to eliminate a few typographical and grammatical errors. For example, "interpet" should be corrected to "interpret" (p. 3), and "engeneering" should be corrected to "engineering" (p. 9).Typographical and grammatical corrections
We thank the referee for noting these errors. We will carefully proofread the revised manuscript and correct the identified words, together with any additional typographical and grammatical errors.Citation: https://doi.org/10.5194/egusphere-2025-6545-AC3
-
AC1: 'Reply on RC1', Denis Anikiev, 26 Aug 2026
We thank the referee 1 for the detailed and critical assessment. The concerns raised have helped us to rethink both the practical accessibility of the framework and the evidence required to establish its scientific contribution. We agree that the original manuscript did not sufficiently discuss access to alternative LLMs, performance for more complex geological configurations, or the repeatability of the generated geometries. The comment has also helped us to distinguish generation repeatability more carefully from geological uncertainty in a real subsurface interpretation. We have addressed these issues as described below.Reproducibility and access to LLMs
> However, GeoSIRR 1.0 in its current form depends on GPT language models, and an OpenAI API key is required to run the cross-section reconstructions. When attempting to run the provided examples, I received error messages indicating that my API quota had been exceeded and that additional credits would need to be purchased to proceed. This dependency renders GeoSIRR 1.0 effectively inaccessible for reviewing or users without an active paid API subscription, meaning the results cannot be freely tested or reproduced.We thank the referee for this observation and agree that dependence on a paid API was an important practical limitation of the original version. GeoSIRR 1.0.1 is already available through Zenodo:
Anikiev, D., & Mosquera, J. E. (2026). GeoSIRR: Geological Section Interpretation, Reconstruction & Refinement (Version 1.0.1) [Computer software]. https://doi.org/10.5281/zenodo.22097054
as well as GitHub:
https://github.com/CPG-KFUPM/GeoSIRR
In this version, we have extended the framework to support both OpenAI and Ollama backends. Ollama allows a model installed on local hardware to be used without an OpenAI API key or paid OpenAI credits.Moreover, we have deployed GeoSIRR Studio at https://geosirr-studio.onrender.com, which provides a hosted web service for generating and refining cross-sections based on the original GeoSIRR workflow. During the review process, this hosted service is configured so that referees can generate and refine cross-sections without registration or providing their own API key. We will maintain this access until the review process is complete.Pre-generated GeoSIRR text definitions can also be inspected, validated, and rendered using either GeoSIRR 1.0.0 or 1.0.1 without repeating an LLM call. Thus, the geometric results can be examined independently of access to an LLM provider, although reproducing the stochastic generation process still requires LLM calls. In the revised manuscript, we will clarify this distinction and explain that use of a local model does not guarantee the same output quality, runtime, or geological interpretation as another model, because these aspects depend on the selected model, its configuration, and the available hardware.More complex geological scenario
> Furthermore, the manuscript presents only relatively simple tectonic geometries as examples and does not thoroughly demonstrate how the method performs on more complex geometries or tectonic relationships. The authors indicate that increased complexity leads to a significant increase in computation time, yet no further testing is provided for geologically meaningful complex scenarios.We thank the referee for this constructive suggestion and agree that a more geologically involved example would strengthen the manuscript. We plan to add a classical extensional-tectonics example containing several interacting structural elements and sedimentary basin fill. We will use this example to test whether GeoSIRR can translate a more complex geological description into a valid geometry while preserving the relationships among the interacting structures and sedimentary bodies. We will also examine how this increase in geological complexity affects generation time.We will also add a repeated-generation experiment for this scenario and report summary measures of validity, generation time, and run-to-run geometric variation. These experiments will provide evidence of both the framework's representational capability and the aspects of the generated geometry that remain sensitive to stochastic generation.Repeatability and generated-geometry variability
> Moreover, the limitations related to GeoSIRR 1.0, which are noted in the manuscript (no uncertainty quantification, unclear generation performance and repeatability), are serious and the model should not be considered without these issues being properly assessed.We agree that generation performance and repeatability need to be assessed explicitly. In the revised manuscript, we will add repeated-generation experiments using the same model and workflow for a baseline description and a more constrained variant. We will track the validation success, generation effort, and run-to-run geometric variation, supported by numerical summaries and visualizations of the generated geometries. We believe that this comparison will help demonstrate the reproducibility of the workflow, show how prompt specificity affects geometric variability, and clarify that variability among generated models is not equivalent to geological uncertainty in a real subsurface interpretation. We thank the referee for prompting us to provide this evidence.Citation: https://doi.org/10.5194/egusphere-2025-6545-AC1
Model code and software
GeoSIRR: Geological Section Interpretation, Reconstruction & Refinement (1.0.0) D. Anikiev and J. E. Mosquera https://doi.org/10.5281/zenodo.18097054
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 1,192 | 1,096 | 96 | 2,384 | 79 | 113 |
- HTML: 1,192
- PDF: 1,096
- XML: 96
- Total: 2,384
- BibTeX: 79
- EndNote: 113
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Anikiev et al. present GeoSIRR 1.0, a software tool that aims to construct geological cross-sections using Large Language Models (LLMs). Specifically, the authors developed a routine in which qualitative geological descriptions provided as prompts can be transformed into geological cross-sections via LLMs. This method is intended to replace labor-intensive manual cross-section reconstructions.
The code associated with GeoSIRR 1.0 is publicly available through GitHub and Zenodo repositories and the software can be installed easily via the command line. All examples discussed in the manuscript are, in principle, accessible. However, GeoSIRR 1.0 in its current form depends on GPT language models, and an OpenAI API key is required to run the cross-section reconstructions. When attempting to run the provided examples, I received error messages indicating that my API quota had been exceeded and that additional credits would need to be purchased to proceed. This dependency renders GeoSIRR 1.0 effectively inaccessible for reviewing or users without an active paid API subscription, meaning the results cannot be freely tested or reproduced.
Furthermore, the manuscript presents only relatively simple tectonic geometries as examples and does not thoroughly demonstrate how the method performs on more complex geometries or tectonic relationships. The authors indicate that increased complexity leads to a significant increase in computation time, yet no further testing is provided for geologically meaningful complex scenarios. Moreover, the limitations related to GeoSIRR 1.0, which are noted in the manuscript (no uncertainty quantification, unclear generation performance and repeatability), are serious and the model should not be considered without these issues being properly assessed.
For the reasons outlined above, I do not consider the manuscript to demonstrate a sufficiently innovative or significant scientific contribution while its results remain non-reproducible in its current form. Therefore, I recommend against acceptance.